Semantic segmentation algorithm for remote sensing images based on PSPNet

Remote sensing images contain various object categories, with complex backgrounds and diverse object scales. Single-scale feature extraction methods struggle to capture both large-scale and fine details simultaneously, thus influencing the accuracy of segmentation. Therefore, in this study, we employed the PSPNet model for segmentation due to its Pyramid Pooling Module (PPM) that can extract multi-scale features. To improve the model's capacity for global information interaction and representation, we used Swin Transformer to replace ResNet-101 as the backbone of PSPNet. Ultimately, our model achieved a mean Intersection over Union (mIoU) of 52.1%, which represents an improvement of 1.32% compared to PSPNet with ResNet-101 as the backbone. Under the same conditions, We also performed model comparisons between DeepLabV3+ and SETR. demonstrating the advantages of our improved model.